Self-supervised learning (SSL) has shown great promise in computer vision, enabling networks to learn meaningful representations from large unlabeled datasets. SSL methods fall into two main categories: instance discrimination and image modeling. While instance discrimination is fundamental to SSL, it was originally designed for classification and may be less effective for downstream tasks that require fine-grained or spatially localized representations. In this focused survey, we study SSL for object detection under challenging practical conditions, with particular emphasis on small object detection, domain shift and few-shot learning. Building upon previous surveys, we not only provide a detailed comparison of SSL strategies, but also assess their effectiveness for object detection using both CNN and ViT-based architectures. Our benchmark is performed fairly by fine-tuning a Faster R-CNN initialized with several exemplary SSL methods ourselves, including object-level Instance Discrimination and Masked Image Modeling methods, on the widely used COCO dataset, as well as on a domain-specific dataset focused on vehicle detection in infrared remote sensing imagery. We also evaluate the impact of pre-training on custom domain-specific datasets, highlighting how some SSL strategies are better suited for handling uncurated data. Furthermore, we assess the methods in few-shot settings and inference on noisy input, revealing important behavioral differences depending on the type of encoder used. Our findings highlight that combining approaches with complementary local and global biases improves performance across the evaluated object detection settings. Overall, this survey provides a practical guide for selecting optimal SSL strategies in different scenarios. • We propose a survey on self-supervised learning for real-world object detection. • In our benchmarks, we pay attention to small object detection performance. • Challenging conditions such as frugal setting or remote sensing data are considered. • The benefits of pre-training on custom domain-specific datasets is assessed. • A road map for selecting appropriate self-supervised learning strategies is provided.

Infrared Small Target Detection (IRSTD) is a challenging task in defense applications, where complex backgrounds and tiny target sizes often result in numerous false alarms using conventional object detectors. To overcome this limitation, we propose an Anomaly-Aware version of You Only Look Once (YOLO) detector (namedAA-YOLO), which integrates a statistical anomaly detection test into its detection head. By treating small targets as unexpected patterns against the background, AA-YOLO effectively controls the false alarm rate. Our approach not only achieves competitive performance on several IRSTD benchmarks, but also demonstrates remarkable robustness in scenarios with limited training data, noise, and domain shifts. Furthermore, since only the detection head is modified, our design is highly generic and has been successfully applied across various YOLO backbones, including lightweight models. It also provides promising results when integrated into an instance segmentation YOLO. This versatility makes AA-YOLO an attractive solution for real-world deployments where resources are constrained. The code is available at https://github.com/AMIAD-Research/AA-YOLO . • We present AA-YOLO, a YOLO model with anomaly testing for IR small target detection. • AA-YOLO attains state-of-the-art detection of small targets in complex IR scenes. • Our method is robust across varied scenarios: limited data, noise, and domain shifts. • AA-YOLO is lightweight and works with many YOLO backbones, ideal for real-world use.

State-of-the-art red-green–blue (RGB) texture synthesis algorithms rely on style distances that are computed through statistics of deep features. These deep features are extracted by classification neural networks that have been trained on large datasets of RGB images. Extending such synthesis methods to multispectral images is not straightforward, since the pretrained networks are designed for and have been trained on RGB images. In this work, we propose two solutions to extend these methods to multispectral imaging (MSI). Neither of them requires additional training of the neural network from which the second-order neural statistics are extracted. The first one involves optimizing over batches of random triplets of spectral bands during training. The second one projects multispectral pixels onto a 3-D space. We further explore the benefit of a color transfer operation upstream of the projection to avoid the potentially abnormal color distributions induced by the projection. Our experiments compare the performances of the various methods through different metrics. We demonstrate that they can be used to perform exemplar-based texture synthesis, achieve good visual quality, and come close to state-of-the-art methods on RGB bands. Code is available athttps://github.com/selim2483/multispectraltextureCNN.

The paper deals with the detection performance of a moving target in multispectral IR image sequences with lowSNR. In this context, track-before-detect (TBD) is the generally used method, which consists in accumulating raw images over time to track before detect. For each target hypothesis (position, velocity, amplitude), the signal is integrated over time. In this way, the potential target with the best spatio-temporal correlation will be a candidate for a detection test. In this paper, we develop a minimum detection bound regardless of the TBD method used. Specifically, this bound gives the minimum average number of multispectral images required to detect the target.

Detecting small to tiny targets in infrared images is a challenging task in computer vision, especially when it comes to differentiating these targets from noisy or textured backgrounds. Traditional object detection methods such as YOLO struggle to detect tiny objects compared to segmentation neural networks, resulting in weaker performance when detecting small targets. To reduce the number of false alarms while maintaining a high detection rate, we introduce an a contrario decision criterion into the training of a YOLO detector. The latter takes advantage of the unexpectedness of small targets to discriminate them from complex backgrounds. Adding this statistical criterion to a YOLOv7-tiny bridges the performance gap between state-of-the-art segmentation methods for infrared small target detection and object detection networks. It also significantly increases the robustness of YOLO towards few-shot settings.

The detection of tiny objects is a challenging task in computer vision. Conventional object detection methods have difficulties in finding the balance between high detection rate and low false alarm rate. In the literature, some methods have addressed this issue by enhancing the feature map responses for small objects, but without guaranteeing robustness with respect to the number of false alarms induced by background elements. To tackle this problem, we introduce an a contrario decision criterion into the learning process to take into account the unexpectedness of tiny objects. This statistic criterion enhances the feature map responses while controlling the number of false alarms (NFA) and can be integrated as an add-on into any semantic segmentation neural network. Our add-on NFA module not only allows us to obtain competitive results for small target, road crack and ship detection tasks respectively, but also leads to more robust and interpretable results.

Deep learning based methods for single-image super-resolution (SR) have drawn a lot of attention lately. In particular, various papers have shown that the learning stage can be performed on a single image, resulting in the so-called internal approaches. The SinGAN method is one of these contributions, where the distribution of image patches is learnt on the image at hand and propagated at finer scales. Now, there are situations where some statistical a priori can be assumed for the final image. In particular, many natural phenomena yield images having power law Fourier spectrum, such as clouds and other texture images. In this work, we show how such a priori information can be integrated into an internal super-resolution approach, by constraining the learned up-sampling procedure of SinGAN. We consider various types of constraints, related to the Fourier power spectrum, the color histograms and the consistency of the upsampling scheme. We demonstrate on various experiments that these constraints are indeed satisfied, but also that some perceptual quality measures can be improved by the proposed approach.

Existing computer simulations of aircraft infrared signature (IRS) do not account for dispersion induced by uncertainty on input data, such as aircraft aspect angles and meteorological conditions. As a result, they are of little use to estimate the detection performance of IR optronic systems; in this case, the scenario encompasses a lot of possible situations that must be indeed addressed, but cannot be singly simulated. In this paper, we focus on low-resolution infrared sensors and we propose a methodological approach for predicting simulated IRS dispersion of poorly known aircraft and performing aircraft detection on the resulting set of low-resolution infrared images. It is based on a sensitivity analysis, which identifies inputs that have negligible influence on the computed IRS and can be set at a constant value, on a quasi-Monte Carlo survey of the code output dispersion, and on a new detection test taking advantage of level sets estimation. This method is illustrated in a typical scenario, i.e., a daylight air-to-ground full-frontal attack by a generic combat aircraft flying at low altitude, over a database of 90,000 simulated aircraft images. Assuming a white noise or a fractional Brownian background model, detection performances are very promising.

We propose a new method, based on level sets, to detect aircraft on low resolution infrared images. Aircraft correspond to hot temperatures at the sensor level. Hence it is natural to rely on a test that considers the hottest pixels in the sensed image. If these pixels are close, they are likely to come from a target (i.e., an aircraft); otherwise they belong to the clutter. Instead of manually testing the neighborhood of each hot pixel, we use level sets; this is the first contribution of the paper (corresponding to eq. 2). The other contribution is the calibration of the resulting test. The method is implemented and tested over a database containing 45 604 simulated aircraft images and provides 98.5% of correct detections.